The world loves gambling. It’s one of the few things on earth that humanity seems to get behind everywhere we go. The glitz of Monaco’s Monte Carlo1, the mighty riverboats of the Mississippi2, and the millions of lights of Macau3 all glisten with gamblers’ gold. In New South Wales alone, more than $1 Million AUD is lost to pokies every single hour.4. $2.3 Billion over the course of 3 months. Just people captivated by the poker machine, sitting down and dutifully feeding it cash.
One person in particular has been blessed by the lottery gods. Elon Musk is officially the world’s first trillionaire, after the float of SpaceX5. Fortune certainly favours the bold (primarily with Canadian passports6 and emerald mine money7).
SpaceX is certainly an interesting company. Famous not only for its rockets, but for Starlink, a genuinely impressive service that provides affordable, reliable satellite internet service to parts of the world that have traditionally lacked it. As someone with a boyfriend currently in rural Western Australia, it’s been the only way I’ve been able to regularly keep in close contact with him, and I’m deeply appreciative for it.
But not all that glistens is gold. Not even half a year ago, Musk decided to merge SpaceX with his Artificial Intelligence company, xAI8. While, true, SpaceX does have a digital DNA in its custody of Starlink, I’m not convinced that the business entities were ever similar enough to benefit from integration. Not my circus, not my monkeys, though. Which I’m thankful for, because it looks like the billions of dollars that have been tipped into A.I. will never make economic sense9.
A big part of the problem here is that A.I., in the form of Large Language Models (LLMs), is, itself, a giant gamble in that the output is probabilistic, not deterministic. When something is deterministic, rerunning a task with the exact same inputs yields the same result every time. When I turn the key in my car’s ignition, it’s deterministic, so the same key starts the same car the same way. Compare this something where there’s a chance something may or may not happen - that’s probabilistic. If you take an LLM and ask it the exact same question 5 times, you’ll get 5 different answers. It’s a statistical guesser, not a reasoned problem solver10. So when you ask an LLM to generate work without flaws, it’s really not far removed from asking a poker machine to generate winnings without losses.
In terms of a business proposition for A.I, this is a massive fly in the ointment. Business processes tend to be pretty deterministic; business goals, laws and regulations, leadership drives, and an assortment of other ingredients turn into processes, policies, and ways of working. Is this deterministic? In theory, it is. If you put it through a traditionally programmed computer system, it certainly can be (more or less). If a human does it? In fairness, no. People are people, and human error at the very least introduces randomness into whatever it is you’re trying to achieve.
If the probabilistic nature of human activity and the probabilistic nature of LLMs were approximately the same, then there wouldn’t be an issue. Thing is though, they aren’t. When it comes to human interaction, there is an intrinsic transparency and accountability that comes with that. Management isn’t perfect either, but there are well understood tools in the toolbox to help ensure human alignment to the broader business goals at play. We can see what people are doing. We can ask about what they’ve done, and why, and what motivations there were. We can come to common understandings, and overlay broader goals on everyday actions.
But A.I. cannot do that. The lack of critical thinking means its probabilistic nature will always be a gamble. If you ask it to write code, to construct an essay, to assess text, there are no human-like heuristics in the mix. You cannot trust that it implicitly understands what you want it to do. There is no understanding. Tokens churning through a CPU don’t have any more intuition than bingo balls do. But it’ll create something that looks correct, and it’ll present it with such authority and confidence that of course it knows what it’s talking about.
This presents a dilemma to business. Ignore the cost of tokens. Ignore the data security issues. Focussing on A.I. outputs, they were all 100% statistical probabilities. They might be correct. They might not be. How are you to tell? If the point of introducing A.I. to business functions was to reduce human input, then what’s the point in hiring all the same humans to check the work? You could, of course, get another A.I. to check the work, but you haven’t actually solved the root issue; your assessment A.I. has all the same intrinsic afflictions.
So with invisible errors being largely inevitable, what’s the impact? Depends, really; an on-the-fly translation service has a very different risk profile to one that generates code that underbeds financial, healthcare, or other sensitive data. Meta found that out the hard way, when its A.I. support bot took a pretty loose view on account security (or, according to Fox, Obama revealed himself to be the pro-Iran fanatic they always knew)11.
Businesses, of course, can’t ignore token cost. It’s a cost that’s rapidly adding up, too; in some cases, A.I. token usage is costing more than humans are12. If businesses aren’t saving money on output generation, and then have to spend more on output quality checks, is the big gamble on A.I. going to pay off? I can think of one trillionaire who’s got a lot riding on it.
